PHYSICS-INFORMED NEURAL NETWORKS FOR MARTIAN THERMAL DIFFUSIVITY RECOVERY AND THE ROLE OF THE MEASUREMENT MODEL

We apply physics-informed neural networks (PINNs) to recover the thermal diffusivity of Martiansubsurface soil from NASA InSight HP3 mission data, the first in-situ thermal measurements onMars. A progression of models is developed, from a 1D point-sensor model to a 2D axisymmetricmodel, each addressing a limitation identified in the previous. The point-sensor model produces a440% inconsistency between amplitude-derived and phase-derived κ values, rendering any recoveredparameter physically meaningless regardless of convergence. The spatial-average model recoversκ = 6.28 × 10−8 m2/s with 0.49% convergence spread across three initialisations, confirmed within1% by an independent analytical solution. The 60% discrepancy with Spohn et al.’s value is consistentwith the mole’s thermal fin effect, which the uniform spatial average does not capture. Inverse PINNscan converge reliably to a physically consistent but incorrect parameter value when the forward modeldoes not match the true measurement process. These results demonstrate that for real-instrumentinverse problems, the fidelity of the observation operator matters more than network architecture, lossweighting, or training strategy.